Space research becomes credible when another engineer can inspect the assumptions, run the model, and find the boundary of the claim. A paper alone is not enough. Neither is a code repository without a reference case that shows what the code is intended to test.
Mindpool Labs is the research arm of Mindpool, whose work on on-device artificial intelligenceRunning a model on the device itself instead of sending data to a server. When a network is unavailable, it can be necessary, and it is bounded by the power, memory, and heat the device can spare. meets our robotics pillar directly. On-device models perform inferenceRunning a trained model to get an answer, as opposed to training that model in the first place. On a power-limited device, inference is the operation that has to fit inside the budget. within available power, memory, compute, connectivity, and heat. A rover beyond contact and a servicing robotRepairing, refueling, inspecting or moving a satellite while it is still in orbit, instead of replacing it. deciding during eclipseThe stretch of an orbit spent inside a body's shadow, where no sunlight reaches the solar arrays. The vehicle runs on battery, and its thermal balance inverts from shedding heat to holding it. face related limits when planning actions, which is why this laboratory studies autonomy that carries energy and thermal state into planning. Labs has an open-publication policy for generic research models, benchmarksA fixed set of scenarios, metrics and baseline methods that anyone can run, so that results from different groups can be compared honestly., papers, and software. Mindpool may carry selected work into device and product integration. Partner-specific parameters and vehicle-specific implementations remain private, while public reference artifacts will be released under their stated licenses.
Why a software and physics team belongs in space
The space economy is usually described through its platforms: rockets, satellites, landers, and stations. They are the most capital-intensive layer of the stack, which is one reason they receive most of the attention. Beneath them sits the science, while software runs across every layer — simulation, planning, digital twinsA simulation of one specific physical vehicle, kept in step with that vehicle while it operates. It carries the as-built parameters and measured wear of that individual unit, fed by its telemetry, so it predicts what the real thing will do rather than what a design of its type would do., and the physics models that make the larger system trustworthy. A team drawn from software, data, artificial intelligence, and academic physics can produce a visible proof in that horizontal band without first owning a test stand. It is where we begin.
How the work earns confidence
Each initiative begins with a bounded reference missionA bounded, published mission scenario used to test one research claim. It fixes the assumptions, initial conditions, limits, metrics and comparison method so that another team can run the same study., a named baselineThe existing or simpler method against which a new result is compared. Both methods must receive the same hardware assumptions, starting conditions and environment for the comparison to mean anything., and limits that can be checked. The specifications, model, and benchmark develop together because each exposes assumptions the others must carry. We intend to test the first result across declared uncertainty and seek an independent rerun before hardware work begins. Hardware enters when the analytical result is stable enough to justify a test at a partner facility.
Two pillars, one umbrella
Space is the umbrella. Beneath it we have chosen two pillars where the unsolved problems are plentiful and the physics is central. Energy covers generation, transmission, storage, and above all heat rejectionGetting waste heat out of a vehicle. In space, a radiator panel emits that heat as radiation. Heat rejection can constrain high-power spacecraft before available electrical power does.. Robotics covers in-orbit servicing, surface operations, manipulation, and autonomy under communication delay. Our first initiative sits exactly where the two meet.
Principles
Open by default
Publishable code will be released under Apache-2.0A permissive open-source license. It allows commercial use, modification and redistribution, and it carries an explicit patent grant., with documents and data under Creative CommonsA license that allows documents and data to be reused, including commercially, as long as the original is credited.. Decisions that affect a public model will be recorded in design documents.
Papers with artifacts
A paper will not be published without runnable code, stated inputs, and the baseline used for comparison.
Partners before model freeze
Thermal, autonomy, and scenario reviewers enter while assumptions can still change. Chambers, benches, and test stands come later, from laboratories that already operate them.
Bound every claim
Results will state their operating conditions, uncertainty, constraint violationsAn occasion when a planned or simulated state crosses a declared limit, such as a battery falling below its minimum charge or a component exceeding its allowed temperature. Reporting both the count and size of these events prevents useful-work metrics from hiding an unsafe plan., and known failure modes. A useful negative result advances the work when it shows where the method stops.
Contact
Write to research@mindpoollabs.com to discuss a paper, a benchmark scenario, or a partnership.